Introduction

Stakeholder

• Worked with W. Evan Johnson of Rutgers University

Mission

• Deployed bioinformatic models to analyze diseases using genomic expression data

• Focussed on CVID: a clinically defined disease, but which is molecularly not well-defined

• Organized a complete code-base to support immunological research and publishability

Context

This analysis is subsequent to work of Dr. Paul Maglione : ‘Convergence of cytokine dysregulation and antibody deficiency in common variable immunodeficiency with inflammatory complications’ :

• Preceding study involved LPS stimulation of peripheral blood tissue and identified “potential significance of reduced LPS-specific antibodies”… “suggesting a a 2-hit model of pathogenesis requiring further exploration”

Local and Global Impact

We hope our analysis will contribute as support system and a viewpoint on :

Continuing Professional Development

• Jaccard Coefficient for similarity analysis
• Permutation Tests for analysis design validation
• Bioinformatics pipelines for Differential Expression Analysis and Pathway Analysis (Summarized Experiment, DESeq2, Limma, and enrichR Libraries)
• Introductory immunology and clinical practices
• Professional academic journal writing

Data Science Lifecycle

1. Data acquisition and representativeness

Dr. Johnson provided three separate data sets directly. These were developed or measured by research teams at Boston University and Rutgers:

1. Nasopharynx sampling of controls and CVID-diagnosed patients
2. Peripheral Blood Mononuclear Cell samples (white blood cells found in bloodstream) for controls and CVID
3. Nasal and Blood samples from patients living with Tuberculosis and controls

As with other genomic studies, this analysis was subject to the problem of “small N big P”

2. Data management

• We learned and used the Summarized Experiment (S4) data structure. This structure packages count data with patient metadata into a single object.

• Patient anonymity was a key priority and maintained throughout.

• Maintained separate databases while sharing versions through Git to prevent leaking sensitive data

3. Data preparation and integration

• General data cleaning (renaming variables, filtering outliers, splitting datasets)

• Augmented the data by converting gene expression counts into: counts per million (CPM), log-counts, and log-CPM

• Moderate EDA performed to identify and label appropriate treatment groups

4. Data analysis

• Dimensionality reduction by PCA (axial) and UMAP (maintaining distances)

• DESeq2 model deployed to identify significant genes. DESeq2 uses negative binomial regression to model count data, which estimates variance and accounts for over-dispersion. For gene \(i\) in sample \(j\), the observed count \(K_{ij}\) follows a negative binomial distribution with mean \(\mu_{ij}\) and gene-specific dispersion parameter \(\alpha_i\):

\[ K_{ij} \thicksim NB(\mu_{ij}, \alpha_i) \]

• Pathway analysis using enrichR provided databases to cross-reference the statistically significant genes against known cellular pathways (collection of genes). This enables interpretation of DESeq results using curated biological knowledge.

• Comparative analysis using Jaccard coefficient statistic and permutation tests to validate the design of our analysis

5. Model development/deployment

Model development wasn’t a central concern of this project. While we used DESeq models, they were used for inference.

Theoretical model development might involve:

• Leveraging the significant genes to create a classification model to predict whether a patient has CVID or not

• Choosing among the multitude of classification models (Naive Bayes seems like a good candidate)

• Training and testing the model

• Evaluating the model using accuracy, confusion matrix (and its derivatives), and Cohen’s Kappa

6. Communication of Knowledge Obtained from Data

• Creation of subject appropriate statistical plots : UMAP plots, Heatmaps, Volcano plots

• Adjusting figures for publication-readiness

• Rslidy for clear and concise presentation, transmission of results

• Journal-appropriate writing, in particular for Methods section

Non-technical Discussion

Legal, professional, ethical, security, and social issues were discussed in depth:

1. Samples are described with metadata for statistical analysis, but must be separately analyzed for discretion around patient anonymity.
2. The social ethics of designations and sensitive usage were discussed: samples taken in Uganda involved patients living with HIV. Consideration of care and personal ethics and sensitivies is important in designating these subjects. Diagnosis of Tuberculosis may warrant similar sensitivities.
3. Presently evolving landscape of clinical research and public funding were discussed with Dr. Johnson, including professional practices and current risks

Development Methodology

Outside Classes

We relied on many courses central to and supportive of the Data Science degree track:

Conclusion

Thank you